AI agents in GTM are in a confusing spot right now. A lot of companies are pushing them, and almost no one can tell you what they're actually for. Should you build copilots for your team? Autonomous systems that run on their own? Replace the whole department with AI and call it a day?
Everyone has an opinion. Most of the loud ones are selling something.
But before any of that, there's a more basic question: do you need AI agents in the first place?
My answer, for most of GTM, is no — and the place agents do belong is probably not where you've been told to put them. Let me walk through why.
Four words that get used interchangeably (and shouldn't)
Most of the confusion comes from the fact that "workflow," "AI workflow," "agent," and "autonomous system" get thrown around as if they mean the same thing. They don't, and the difference is the whole game.
A workflow is a fixed set of steps you design up front. Step one, then two, then three. Same path every time. A sequence in your sales tool, a Zapier automation — that's a workflow. It's dumb in the good sense: it does exactly what you told it, nothing more.
An AI workflow is the same fixed path, except AI now handles the steps that used to need a human brain — writing the email, classifying the ticket, summarizing the call. You still design the path. The AI just fills in the judgment-heavy slots inside it.
An AI agent is different in kind, not degree. You don't give it a path. You give it a goal. It decides the steps itself — what to do first, which tool to reach for, when it's done. It loops until it thinks it's finished.
An autonomous system is agents and workflows wired together to run an entire process end to end, with no human in the loop.
The jump from a plain workflow to an AI workflow is enormous — you go from rigid automation to something that can read, write, and reason. The jump from an AI workflow to an agent? Much smaller than the hype suggests. You're mostly trading one thing for another. And whether that trade is good depends entirely on what you're using it for.
The AI SDR, step by step
Take the most hyped GTM example: the AI SDR. What does it do? The same thing a human SDR does — research the person, map them to the pitch, write the email, send it. Yes, I'm simplifying, but the point holds: every time you write that email, you run the same type and number of steps, start to finish.
That's a workflow. A good one, maybe even one with AI doing the writing — but a workflow.
The catch with workflows — and honestly, I'd call it a feature more than a downside — is that you have to design them. Start to finish. You think through the edge cases, the fallbacks, the if-this-then-that branches. That takes time and it takes expertise. But in exchange you get something precious: you know exactly what the system will do before it does it.
Now picture the agentic version of the same SDR. You don't hand it the steps. You hand it a goal — "book meetings with our ICP" — and let it figure out the rest. It decides which 200 of your 5,000 leads are worth contacting. It picks the channel per person. It chooses when to follow up, how many times, and when to give up. It might decide to skip someone you'd have emailed, or hammer someone you'd have left alone.
That's the promise of agents: you don't design any of it. The agent works it out. You give up consistency to gain flexibility.
Sounds great. But do you actually need that flexibility in GTM?
Most of GTM is customer-facing — and that changes everything
Look at what GTM work actually is. Emails. Content. Sales and marketing materials. Customer tickets. The proposal, the follow-up, the reply to the upset account. Almost every piece of it is customer-facing — every small interaction that builds credibility, trust, loyalty, and, eventually, revenue.
And here's the problem with handing that to an agent: an agent has no accountability. It has no concept of doing something wrong. It will keep doing exactly what it's pointed at, even when what it's doing is damaging your business — and it will do it with total confidence.
"Fine," you say, "I'll keep an eye on it."
Will you, though? Can you read the thousand emails it sent yesterday? The thousand it'll send today? Maybe you build a second agent to watch the first one — but then, who watches the watchmen?
This is the real issue, and it's worth being precise about it. The danger isn't that an agent makes mistakes. Every system makes mistakes. The danger is the asymmetry: customer-facing mistakes compound silently. A workflow that breaks usually breaks loudly — it errors out, it stops, you notice. An agent that's subtly wrong just keeps going, politely, at scale. You don't find out from a log. You find out from churn, from a deal that went cold for no reason you can name, from a prospect who quietly decided you weren't serious. By the time the data tells you, the damage is already booked.
So here's the first rule of the day: everything customer-facing should be an AI workflow. Not an agent. You want the consistency, and you want to be able to point at any step and know why it happened.
Why agents work for code and not for customers
People look at how well agents work in software engineering and assume it'll transfer. It won't, and the reason is instructive.
Code is hidden. It lives in the codebase. Sure, there are purists who care how cleanly it's written, but for the most part: if it works, it ships, and you can always come back and make it better later. The defining property of code is that it's reversible. A bad function can be rewritten before or after it ships. A messy module can be refactored next sprint.
A sent email is not reversible. You can't come back three months later, resend the follow-up, and act as if nothing happened. The customer already read the first one. The interaction already landed.
That's the real distinction — not "stakes," because plenty rides on bad code too: shipped bugs, security holes, outages. The distinction is reversibility. When a coding agent gets it wrong, you usually get to fix it. When your GTM agent gets it wrong, the moment is gone, and sometimes the deal goes with it.
Where agents absolutely earn their place
None of this means you shouldn't use agents. It means you should point them at the right work.
A huge amount of what your team does never reaches a client or a prospect. Admin work. Coaching. Forecasting. Research. Pipeline hygiene. Internal summaries. The plumbing. For all of that, an agent's flexibility is a gift and its mistakes are cheap — if it gets a forecast slightly wrong or pulls research down a slightly odd path, you catch it internally, before it ever touches a customer.
So the second rule writes itself: agents for internal work. The exact place customer-facing logic should be a rigid workflow is the place internal work should be free to roam.
Autonomous systems: the powerful, dangerous third option
Then there are autonomous systems — agents and workflows fused into something that runs a whole process on its own. These are powerful. They're also the easiest way to hurt yourself, because by definition you've taken your hands off the wheel. You need to know what you're doing, why, and how, very clearly, before you commit.
There are really only two situations where they make sense, and they look like opposites until you see what they share.
One: work that was never being done at all. Here's an example. You're a SaaS company with a mountain of free users. Your sales team has never had the capacity to engage or upsell them, and realistically never will. So that revenue just sits there, untouched, leaking out of the business every month. There's no existing standard to violate, no relationships to damage, no downside to "doing it worse than before" — because before, it wasn't being done. An autonomous system here is pure upside. Worst case, you're back where you started: nobody talking to those users.
Two: work that's been running so long it no longer requires thinking. A process you've operated for years, where you've already hit and handled every edge case, and the day-to-day is mostly copy-paste with the brain switched off. You're not asking the system to figure anything out — you've already figured it out. You're just handing off the execution of something fully mapped.
In both cases the judgment is settled. In the first, there's no standard to fall short of. In the second, the standard is already encoded. Autonomous systems fail when you ask them to invent judgment on the fly in front of customers — and that's exactly what these two cases don't require.
So, do you need agents in GTM?
You need all three. The trick is matching the tool to the work:
- AI workflows for anything customer-facing — where consistency and accountability matter more than flexibility, and where a mistake can't be taken back.
- AI agents for internal work — admin, research, forecasting, coaching — where flexibility pays off and mistakes stay in-house.
- Autonomous systems for work that was never being done in the first place, or work so routine it no longer requires thinking — the two cases where judgment is already settled.
The hype tells you to ask "agent or no agent?" That's the wrong question. The right one is "is this in front of a customer or not?" Answer that, and the rest of the decision makes itself.